The money underneath the AI build-out came into view this week. Nvidia agreed to guarantee $105 billion so a single customer could sign a single lease, the Journal counted roughly $3 trillion in commitments that never appear as debt, and Anthropic offered $6 billion for a company whose product is making chips work harder. Meanwhile OpenAI published a study of its own enterprise customers that couldn't find a link between using AI and earning more per employee, China quietly won the open-model download race, and a supply chain breach handed out the credentials of 2,500 companies.
1. Nvidia is guaranteeing $105 billion of OpenAI's rent
Nvidia disclosed in a securities filing on Monday that it will guarantee up to $105 billion in lease and power obligations so OpenAI can take a 20-year lease on a 10-gigawatt data centre campus in Ohio built by SoftBank's SB Energy, and it is investing a further $1.5 billion in SB Energy itself. The number is down from the $250 billion backstop reported in July, but Jensen Huang said the site alone could contribute $200 billion to Nvidia's revenue and that OpenAI could be worth $600 billion to the company by 2030 — a chipmaker underwriting its largest customer's ability to buy its chips.
2. Big Tech's AI bill is $3 trillion larger than the balance sheets show
The Wall Street Journal reported that roughly $3 trillion in data centre leases, chip purchases and power commitments sit outside the balance sheets investors actually read, because the obligations are structured as future commitments rather than recorded debt. Nvidia has separately assembled six large asset managers to route $500 billion of third-party capital into data centre projects, which means the leverage underneath the build-out is increasingly held by parties whose exposure is hard to see from the outside.
3. Anthropic is in talks to buy Decart for $6 billion to make Claude cheaper to run
Anthropic is negotiating to acquire the three-year-old Israeli startup Decart for about $6 billion, which would be its largest known acquisition and would fold Decart's team into Anthropic's inference and performance organisation. Decart's software squeezes more throughput out of existing chips, and the size of the offer — a roughly 50% premium to a valuation set three months ago — is a measure of how much the frontier labs will now pay for efficiency rather than capability.
4. Google shipped a new Gemini three weeks after the last one, at half the price
Google released Gemini 3.7 Flash on Thursday, only three weeks after 3.6 Flash, pitching it as its most capable workhorse model for coding and agents and pricing it at 75 cents per million input tokens and $3.75 per million output through the end of the year — half what the previous version cost. The long-promised Gemini 3.5 Pro still has no release date, so the company is competing on cadence and price at the tier businesses actually deploy rather than at the frontier.
5. Alibaba's Qwen passed 3 billion downloads and overtook Meta and Google
Hugging Face data released this week shows Alibaba's open-weight Qwen family has crossed 3 billion downloads and now leads Meta and Google globally, with more than 460 models and 300,000 derivatives built on it, while Chinese open models account for 41% of all open-model downloads worldwide. Meta announced last week that it would open-source its most powerful model and lobby Washington to support the approach; the distribution fight it is entering has largely already been decided.
6. The victim list from the year's biggest AI supply chain breach was published
Security firm CloudSEK released the full impact analysis of the LiteLLM compromise this week: 78,330 secrets taken from the build pipelines of more than 2,500 organisations, including Nvidia, AWS, Samsung, Salesforce, Cisco, ServiceNow, Siemens, FedEx, Volkswagen, HP and the London Stock Exchange Group. The stolen material went well beyond AI provider keys to cloud credentials, SSH keys, Stripe payment keys and npm and Docker publishing tokens, meaning any company whose credentials sat on an affected runner should treat its own downstream software as suspect.
7. Every word Claude writes now carries an invisible watermark
Anthropic confirmed that all models released after August 2 automatically embed a machine-readable watermark in generated text and files, a change driven by European transparency rules that the company is applying worldwide with no option for customers to turn it off. For any business using Claude inside documents, code or customer communications, provenance is no longer a policy question but a property of the output itself.
8. OpenAI's own research found no link between AI use and revenue per employee
Buried in a 69-page OpenAI report on ChatGPT enterprise adoption is the finding that heavier AI use shows no correlation with revenue per employee, the single number most executives would use to justify the spend. Deloitte's survey the same week found 43% of leaders expect agentic AI to significantly disrupt their workforce within 12 to 18 months while half admit they are not investing in the transition, and Salesforce reported agent adoption tripled this year — adoption is real, the return is still being argued about.
9. Coding startups are repricing themselves every three months
Cognition opened talks for a round valuing it at $40 billion less than three months after raising $1 billion at $26 billion, Lovable confirmed a $400 million round at $13.3 billion after closing at $6.6 billion in December, and Higgsfield quadrupled to $5.4 billion in six months. The revaluation cycle for AI application companies is now measured in months rather than years, which is worth knowing if you are signing multi-year contracts with any of them.
10. Most Canadians don't want public money going to AI data centres
A Nanos poll published this week found a majority of Canadians oppose government support for AI data centres, citing electricity demand, environmental harm and water use, days after hundreds of Albertans packed a county office in Morinville to protest Meta's proposed one-gigawatt facility. The same week Nvidia guaranteed $105 billion for a single campus in Ohio, the social licence for building them in Canada is visibly narrowing.
